The Shift to Probabilistic Design

Traditional design models, such as those proposed by Jesse James Garrett and Jamie Mill, focused on deterministic outcomes where designers mapped every user action to a specific interface state. Generative AI breaks this model because the underlying system is probabilistic, creating emergent behaviors that cannot be fully predicted. Designers must now act as 'full-stack' practitioners—not by becoming engineers, but by becoming multilingual in the layers that sit beneath the interface.

The Four Layers of AI UX

AI experiences are composed of four interdependent layers. Designers must understand how each influences the others to create products that are legible, safe, and valuable.

1. The User Interface Layer

Interfaces are shifting from primary drivers of interaction to oversight tools. Early in a journey, the UI facilitates direct instruction (prompts, workflows). As the system builds context, the UI's role shifts toward monitoring and orchestration. Designers should move away from 'chat-only' thinking toward a palette of inputs—inline actions, ambient nudges, and persistent controls—that match the user's current level of trust and system autonomy.

2. The Context Layer

This layer acts as the engine of the experience, storing user intent, preferences, and historical data. It prevents the 'cold start' problem. Designers must practice 'context engineering'—deliberately managing what information is passed to the model. The challenge is balancing enough context for personalization against 'context rot,' where excessive or irrelevant data degrades model performance.

3. The Harness Layer

This is the operational layer that allows agents to act independently. It includes:

  • Connectors: Rules for data access and third-party integrations.
  • Tools: Specific capabilities (e.g., writing emails, updating records) that models use to execute tasks.
  • Skills: Reusable working knowledge and methods for processing information.
  • Agents: Autonomous systems that coordinate tools and skills to achieve goals. Designers must ensure these background processes are visible and controllable, allowing users to delegate tasks without losing the ability to intervene.

4. The Model Layer

Models are not interchangeable commodities; they have distinct 'personalities' based on their training data and tuning. Designers must choose models based on specific task requirements—such as reasoning depth versus speed—and configure their behaviors (e.g., verbosity, confidence, latency) to match the product's needs. The goal is to align model capabilities with the user's specific problem space.

Key Takeaways

  • Design for Progressive Autonomy: Start with direct instruction and move toward system-led orchestration as the AI earns the user's trust.
  • Look Below the Surface: Don't just design the chat box; design the context, the tool permissions, and the model behaviors that produce the output.
  • Manage the 'Harness': Treat connectors, tools, and skills as UX components. If users don't understand what an agent can do, they won't trust it to do it.
  • Context is a Map: Treat user context like a video game map—it should be revealed and refined over time, not dumped into the system all at once.
  • Avoid 'Context Rot': Be disciplined about what data is retained. Too much memory can be as detrimental to performance as too little.

Notable Quotes

  • "The work of design looks less like specifying every expected state... and instead closer resembles system design, identifying and manipulating the leverage points in a system that exist in the layers below the surface."
  • "Onboarding into AI products looks less like people learning how to use the system, and more like the system learning how to interpret the user."
  • "The right interface depends on the context surrounding the interaction, like how familiar the user is with the domain, how much the AI knows about them, how sensitive the situation is, and how much confidence the system has in its response."
  • "A good agent experience makes autonomous work feel orchestrated, allowing users to observe and interrupt the model when needed without requiring them to micromanage every step."